KR20230051354A - Method for evaluating malicious comments - Google Patents

Abstract

The present invention relates to a method for determining malicious comments. Comment data is selected by morpheme for each part of speech, and the performance of determining malicious comments can be improved by determining malicious comments only with words corresponding to the five parts of speech, which are nouns, particles, verbs, adjectives, and adverbs, excluding suffixes and punctuation marks, which do not affect the determination of maliciousness. In particular, when extracting words from a sentence, the words are extracted in combination with adjectives rather than extracting nouns alone. When removing verbs, the performance of determining malicious comments can be further improved by building a determination model which removes other parts of speech.

Description

Malicious Comment Detection Method {METHOD FOR EVALUATING MALICIOUS COMMENTS}

The present invention relates to a method for detecting malicious comments, and more particularly, extracts qualities for each part of speech from comments through machine learning and morpheme analysis, filters data by a combination of specific parts of speech, and determines whether a comment is malicious to detect malicious comments. It relates to a malicious comment discrimination method that improves the performance of discrimination.

BACKGROUND OF THE INVENTION [0002] In general, with the dissemination and spread of information and communication technologies such as the Internet, smart phones, and social networking services (SNS), people can easily acquire more information than before. However, as a side effect of the convenience obtained through the development of these technologies, various social problems have occurred due to the overflow of harmful Internet information such as malicious comments, pornography, and fake news.

Among them, malicious comments on the Internet have recently emerged as a social problem, and they have become a hot topic again with the suicide of famous celebrities.

Users who use the Internet do not stop at simply consuming content, but actively express their opinions on the content through comments. In the case of content posted on a web page that is accessed by many users, hundreds or even thousands of comments generated for one piece of content may reach. Public opinion or opinions are formed through comments, and content providers use comments as a major means of receiving user opinions through comments.

However, as the number of comment users increases, the number of users who use comments as a means of expressing their dissatisfaction or maliciously attacking others has also begun to increase. Malicious replies, that is, a new word, 'bad comments', have been coined for comments that exploit the anonymity of the Internet to slander others or spread false information by using profane language such as profanity or obscenity. Some users write comments containing hate expressions, and due to the nature of the Internet, hate comments are exposed to a large number of users, including teenagers, and cause cyberbullying.

There are methods to block malicious comments, but existing methods are focused on filtering out spammy comments. In addition, since existing methods filter hate comments using registered databases, there is a limit to filtering new hate expressions that are not registered in the database.

In addition, according to the conventional comment analysis method, a support vector machine is used to determine whether a word in a comment entered by a user is positive or negative, and the reputation of the comment is analyzed using the determined word to analyze the reputation of the comment. However, since the conventional comment analysis method needs to determine whether all words of the input comments are positive or negative, it takes a lot of time to determine whether the comments are malicious, and it is not efficient. .

Korean Patent Publication No. 10-2009-0103171 (Comment automatic filtering method)

The present invention is intended to solve the problems of the prior art described above, and an object of the present invention is to extract qualities by part of speech from comments through machine learning and morpheme analysis, and filter data by a combination of specific parts of speech to determine malicious comments. It is to provide a malicious comment detection method that can improve performance.

In order to achieve the above object, the malicious comment discrimination method according to the present invention includes a data collection step of collecting comment data using a data collector, extracting morphemes from the collected comment data and pre-processing, and the pre-processing. Selecting morphemes by part-of-speech from the morphemes using a morpheme analyzer; extracting the selected morphemes by part-of-speech as a combination of specific parts of speech; vectorizing the extracted morphemes by part-of-speech; and machine learning by a Support Vector Machine algorithm and determining whether the morpheme includes a word stored in a sentiment type dictionary, and determining whether the comment is malicious through modeling; and It is characterized in that it includes a step of deriving a result value by applying a preset weight and determining whether the comment is malicious if it includes a word stored in .

Here, the parts of speech of the selected morphemes include nouns, particles, verbs, adjectives, and adverbs, and the rest of parts of speech are removed.

Here, the combination of the selected morphemes for each part of speech is characterized in that the combination includes nouns and adjectives and at the same time removes verbs.

Here, the parts of speech are selected using an OKT (Open Korean Text) analyzer, and the vectorization of the morphemes is characterized by vectorization using TF-IDF.

Here, the emotion type dictionary is characterized in that a plurality of emotion dictionaries are constructed in advance according to the type of comments.

Here, the pre-processing operation includes the step of cleansing the content leaving only Hangul from the comment content, the step of removing e-mail, URL, HTML tags and special symbols, and classifying the letters of the comment into consonants and vowels and performing a spell check. It is characterized in that it includes the step of performing a spelling check, the step of correcting the replies according to the spelling, and the step of recombining the corrected replies into sentences.

The malicious comment discrimination method according to the present invention having the above-described configuration sorts the comment data into morphemes for each part of speech, and excludes suffixes or punctuation marks that do not affect the determination of whether or not malicious comments are included. By discriminating malicious comments only with words corresponding to five parts of speech, the malicious comment discrimination performance can be improved.

In addition, the malicious comment discrimination method according to the present invention extracts words from sentences by combining them with adjectives rather than extracting nouns alone, and constructs a discrimination model that removes other parts of speech when verbs are removed to discriminate malicious comments. performance can be further improved.

1 is a flowchart of a malicious comment discrimination method according to the present invention.
2 is a diagram showing an example of word vectorization after nouns and adjectives are extracted and verbs are removed.

Hereinafter, with reference to the accompanying drawings, a malicious comment detection method according to the present invention will be described in detail as an embodiment.

In the method for identifying malicious comments according to the present invention, first, comment data is collected using a data collector (S1).

In the comment data collection step (S1), online comments are collected for various purposes, and in this embodiment, comments included in posts on a portal site are collected using a Python web crawling source. The collection of the comments is carried out by crawling bulletin board and comment information through an application programming interface (API) of portal sites such as Naver and Google.

Then, the replies collected in the comment data collection step (S1) are subjected to behavioral element extraction and preprocessing step (S2).

The pre-processing step includes a step of cleansing the content leaving only Hangul from the comment content (S2-1), a step of removing e-mail, URL, HTML tags and special symbols (S2-2), and converting the letters of the comment to consonants and vowels. , and performs a spell check (S2-3).

An example of the preprocessing step is as follows.

input: Please visit my blog as well. http://blognavercom/whdals0 <h1> Welcome to Seo Yichu </h2> haha

output : Please visit my blog as well. Welcome to Seo Yichu

Then, after performing a spell check, the comments are corrected according to the spelling, the comments are cut based on spacing, the consonants and vowels are separated, and the order of consonants and vowels is combined and compared with a spelling dictionary, Check your spelling.

An example of the step S2-4 is as follows.

input: Please visit my blog as well Welcome Seo Yichu

output: Please visit my / blog / Seo Yichu / Welcome

input: my / blogo / please visit / seo ichu / welcome

output : {ㅅ,ㅔ}, {ㅅ,ㅡ,ㄹ,ㄹ,ㅗ,a,u,ㅇ,ㅔ,c,ㅗ},{ㅅ,아,ㅇ,ㅁ,ㅜ,ㄴ,ㅎ,ㅐ,ㄴ ,TT,ㅅ,ㅔ,ㅇ,ㅛ}

output : {ㅅ,ㅔ}, {ㅅ,ㅡ,ㄹ,ㄹ,ㅗ,a,ㅡ,ㅇ,ㅔ,c,ㅗ},{ㅅ,아,ㅇ,ㅁ,ㅜ,ㄴ,ㅎ,ㅐ,ㄴ ,TT,ㅅ,ㅔ,ㅇ,ㅛ} (correct spelling compared to spelling dictionary)

Then, a step (S2-5) of recombining the modified comments into sentences is performed.

input : {ㅅ,ㅔ}, {ㅅ,ㅡ,ㄹ,ㄹ,ㅗ,a,ㅡ,ㅇ,ㅔ,c,ㅗ},{ㄴ,아,ㅇ,ㅁ,ㅜ,ㄴ,ㅎ,ㅐ,ㄴ ,TT,ㅅ,ㅔ,ㅇ,ㅛ}

output: Please visit my blog as well

Then, morphemes are selected according to parts of speech from the preprocessed morphemes using a morpheme analyzer (S3).

In this embodiment, analysis is performed through a morpheme-based feature extraction method to determine malicious comments. In this specification, 'feature' extraction refers to processing data composed of sentences or words into a data form usable in a machine learning model.

In this embodiment, the morpheme analyzer uses OKT (Open Korean Text) analyzer provided by KoNLPy. Using the OKT morpheme analyzer, 8 morphemes with a high frequency that can affect the determination of malicious comments are selected from 28 extractable morphemes.

In this embodiment, the frequency of appearance of morphemes was checked in 10,000 malicious comments, 10,000 non-malicious comments, and 10,000 mixed malicious and non-malicious comments, and words were extracted based on the corresponding morphemes and used as qualities, and malicious comments [Table 1] shows the frequency of each 10,000 morphemes.

Frequency of occurrence by morpheme of 10,000 malicious comments morpheme number of appearances frequency ranking Noun 238,641 One Josa 91,709 2 Verb 71,971 3 Punctuation 30,401 4 Adjective 28,831 5 Foreign 23,834 6 Suffix 15,718 7 Adverb 9,499 8 Number 7,126 9 KoreanParticle 6,796 10 Determiner 4,385 11 Alpha 3,905 12 Connection 1,009 13 Exclamation 820 14 URL 560 15 ScreenName 281 16 eomi 200 17 HashTag 45 18 Email 2 19 Unknown 0 20 Proeomi 0 20

As a result of the experiment, more than 10,000 nouns, particles, verbs, punctuation marks, adjectives, foreign words, and suffixes appeared in malicious comments, and more than 10,000 nouns, particles, verbs, adjectives, punctuation marks, and suffixes appeared in non-malicious comments. In the mixed data of comments and non-malicious comments, there were more than 10,000 nouns, particles, verbs, punctuation marks, adjectives, suffixes, foreign words, and adverbs.

In this embodiment, among a plurality of morphemes with high occurrence frequency, punctuation marks, suffixes, and data that are not useful for classification in the case of foreign languages occupy most of the data, so five morphemes such as nouns, particles, verbs, adjectives, and adverbs are excluded. Include only (part of speech) and remove the rest of parts of speech.

Then, the selected morphemes for each part of speech are extracted as a combination of specific parts of speech, and the extracted morphemes for each part of speech are vectorized (S4).

Then, the vectorized morpheme is machine-learned by a Support Vector Machine algorithm, and whether the morpheme includes a word stored in the emotional type dictionary is determined to determine whether the comment is malicious through modeling (S5) And, if the comment includes a word stored in the emotion type dictionary, a result value is derived by applying a preset weight, and whether the comment is malicious is determined (S6).

In this embodiment, words extracted based on parts of speech are vectorized using CountVectorizer and TF-IDF, and then malicious comments are identified using SVM.

In this embodiment, CountVectorizer settings are set to analyser = word, min_df = 1.

In this embodiment, the effects of the parts-of-speech-based features on the malicious comment discrimination model were determined through various types of experiments with the extracted features based on five parts of speech.

In the first experiment, 100 qualities were extracted from only words corresponding to the five parts of speech: noun, particle, verb, adjective, and adverb, and an experiment was conducted to determine malicious comments. In the second experiment, after extracting 100 qualities for each part of speech, [ In the third experiment, the maximum number of features was fixed at 500, and the part-of-speech-based features were added at the same rate, and in the fourth experiment, the part-of-speech The experiment was conducted by configuring the features so that words were added according to the frequency of occurrence of the corresponding word. A performance comparison experiment was conducted when the was removed.

The experimental results are shown in [Table 2].

Malicious comment discrimination experiment of parts-of-speech-based quality Experiment number number of qualities Types of parts that perform well (1 is high accuracy) One 2 3 4 5 1-1 100 noun adjective verb adverb inspection 1-2 200 adjective noun verb inspection adverb 2 100 to 500 noun adjective verb adverb inspection 3 500 adjective adverb verb noun inspection 4-1 500 adverb adjective verb inspection noun 1000 verb inspection adjective adverb noun 1500 adjective adverb verb inspection noun 2000 inspection adjective noun adverb verb 2500 noun adjective adverb inspection verb 4-2 500 adjective noun verb adverb inspection 1000 verb inspection adverb noun adjective 1500 adjective inspection verb adverb noun 2000 adjective adverb noun inspection verb 2500 adverb adjective verb inspection noun Experiment number number of qualities Performance ranking when parts of speech are removed 5 500 verb adverb adjective noun inspection

As shown in [Table 2], when extracting words from sentences, higher accuracy was shown when extracting words in combination with adjectives than when extracting only nouns. In addition, it was found that when the verb is removed, the performance is higher than when other parts of speech are removed.

According to the malicious comment discrimination method according to the present invention, as described above, as a combination of morphemes for each selected part of speech, nouns and adjectives are included and verbs are removed at the same time, so that more accuracy is obtained even when a small number of morphemes are extracted. can improve

3 is a diagram showing an example of word vectorization after nouns and adjectives are extracted and verbs are removed.

In the malicious or not malicious determination step (S5), based on the preprocessed data, it is determined whether or not the comment is malicious through machine learning and modeling by a Support Vector Machine algorithm.

SVM is one of the machine learning fields, and when given a set of data belonging to one of two categories, it creates a classification model that determines which category new data belongs to based on the given data set. The created classification model is expressed as a boundary in the space where the data is mapped, and the SVM algorithm is an algorithm that finds the boundary with the largest width among them.

The malignancy determination determines whether the comment includes a word stored in the emotion type dictionary.

The emotion type dictionary is constructed in advance as a plurality of emotion dictionaries according to the type of comments. The emotional dictionary includes hate expression words and profanity words. The profanity dictionary may include, for example, words corresponding to "Korean slang". The emotional dictionary stores new hate expression words discovered by the dictionary construction device and is continuously updated.

In this embodiment, the emotional type dictionary is classified into seven types such as verbal violence, exposure, identity theft, fraud, stalking, bullying, and sexual insult, and is built into each emotional dictionary.

Weights may be applied to words included in each of the sentiment dictionaries, and a separate weight may be set for each word. When the comment includes a word stored in the emotion type dictionary, a preset weight is applied.

Depending on the content of the comment, the weight of the sentiment dictionary may be different. For example, in the case of comments that received weights such as verbal violence 1, exposure 1, sexual insult 1, etc., malicious comments were identified as malicious comments, and each weight was applied to the emotional type dictionary for each type. is extracted, it can be adjusted to block comments by determining them as dangerous malicious comments with 'health risk'.

This embodiment only clearly shows a part of the technical idea included in the present invention, and variations and specific examples that can be easily inferred by those skilled in the art within the scope of the technical idea included in the specification of the present invention It is obvious that all are included in the technical spirit of the present invention.

S 1: Comment data collection step
S 2: pre-processing step
S 3: Morphological Analysis Step
S 4: morpheme vectorization step
S5: malignancy determination step
S 6: Weight application and malignancy determination step

Claims (6)

A data collection step of collecting comment data using a data collector;
Extracting and pre-processing morphemes from the collected comment data;
selecting morphemes according to parts of speech using a morpheme analyzer from the preprocessed morphemes;
extracting the selected morphemes for each part of speech as a combination of specific parts of speech and vectorizing the extracted morphemes for each part of speech;
Machine learning the vectorized morpheme by a Support Vector Machine algorithm, determining whether the morpheme includes a word stored in a sentiment type dictionary, and determining whether the comment is malicious through modeling;
and deriving a result value by applying a preset weight when the comment includes a word stored in a sentiment type dictionary, and determining whether the comment is malicious.
According to claim 1,
The malicious comment discrimination method characterized in that the parts of speech of the selected morphemes include nouns, particles, verbs, adjectives, and adverbs, and the remaining parts of speech are removed.
According to claim 2,
The malicious comment discrimination method, characterized in that the combination of the selected parts of speech by morpheme includes a noun and an adjective and at the same time removes a verb.
According to claim 1,
The parts of speech are selected using an OKT (Open Korean Text) analyzer,
The malicious comment discrimination method, characterized in that the vectorization of the morpheme is vectorized using TF-IDF.
According to claim 1,
The method for determining malicious comments, characterized in that the emotional type dictionary is constructed in advance as a plurality of emotional dictionaries according to the type of comments.
According to claim 1,
In the preprocessing step,
A step of cleansing the content leaving only Hangul from the comment content;
removing e-mail, URL, HTML tags and special symbols;
Classifying the letters of the comment into consonants and vowels and performing a spell check;
After performing a spell check, correcting the comment according to the spelling;
Malicious comment discrimination method comprising the step of recombining the modified comments into sentences

Images